View metadata, citation and similar papers at core.ac.uk brought to you by CORE

provided by Elsevier - Publisher Connector

Permutation Matrices Whose Convex Combinations Are Orthostochastic

Yik-Hoi Au-Yeung and Che-Man Cheng Department of Mathematics University of Hong Kong Hong Kong

Submitted by Richard A. Brualdi

ABSTRACT

Let P,, ., t’,,, be n x n permutation matrices. In this note, we give a simple necessary condition for all convex combinations of P,, . , Pm to be orthostochastic. We show that for n < 15 the condition is also sufficient, but for n > 15 whether the condition is suffkient or not is still open.

1. INTRODUCTION

An n X n doubly stochastic (d.s.) matrix (aij> is said to be orthostochastic (o.s.) if there exists an n X n unitary matrix (a,,> such that aij = IuijJ2. OS. matrices play an important role in the consideration of generalized numerical ranges (for example, see [2]). For some other properties of OS. matrices see [4], [5], and [6]. For n > 3, it is known that there are d.s. matrices which are not 0.s. (for example, see [4]), and for n = 3 there are necessary and sufficient conditions for a d.s. matrix to be O.S. [2], but for n > 3 no such conditions have been obtained. We shall denote by II,, the group of all n x n permuta- tion matrices and by @n the set of all rr X n OS. matrices. Let P,, . . . , P,, E II,. We shall give a necessary condition for conv( P,, . . . , P,,,} c On, where “conv” means “convex hull of,” and then show that the same condition is also sufficient for n < I5.

2. THE CONDITION

The following concept will play an important role.

DEFINITION 1. Let P = (pij) and Q = (qij) be in H,. Then we say that P and Q are cmnpkmentary if, for any I < i, j, h, k < n, pij = phk = qik = 1 implies qhj = I (consequently, qij = qhk = pik = I implies phj = I).

LINEARALGEBRAANDZTSAPPLlCATIONS150:243-253 (1991) 243 0 Elsevier Science Publishing Co., Inc., 1991 655 Avenue of the Americas,New York, NY 10010 0024-3795/91/$3.50 244 YIK-HOI AU-YEUNG AND CHE-MAN CHENG

The following two lemmas are trivial.

LEMMA 1. Let P, Q, R, S E II,. Then P and Q are complementary zyand only ay RPS and RQS are complementary.

LEMMA 2. Let R, S E II,. Then M E q, ayand only afRMS E en.

We shall denote the n X n identity matrix by I,,.

LEMMA 3. L.etP,QEII, and O

Proof. *: Let P = (pii> and Q = (yjj), and let 1 < i,j, h, k ,< n be such that pjj = phk = qik = 1. We may assume i + h and j z k; otherwise, as PEII,, we have i=h and j=k and qhj=qik =l. Suppose that qhjfl. Then qhj = 0. By considering the ith and hth rows of tP +(l - t>Q, we see that tP + (I- t)Q ~5 q,. =: From Lemmas 1 and 2, we may assume P = I,,. Since I, and Q are complementary, Q is symmetric. Hence there is S E ff,, such that

S[tZ,+(l-t)Q]S-‘=J,@ ... @Jk@Zn_2k, where

J~=(,!, ‘L”) for i=l,...,k

and n - 2k 2 0. Consequently, tl, + (1 - t)Q E 0,. n

The following theorem follows immediately from Lemma 3.

THEOREM 1. Let P,, . . . , P,,~rI,.Zfconv(P, ,..., P,n}~@a,thenP, ,..., P, are pairwise complementary.

We are now going to consider whether the condition in Theorem 1 is sufficient. Suppose that P,, . . , P,,L are pair-wise complementary and one of the Pi (i = 1,. . , m> is I,. Then P,, . . . , P,, are symmetric. For symmetric permutation matrices, we have

LEMMA 3. Let P, Q E II, be symmetric. Then P and Q are complemen- tary if and only if PQ is symmetric (i.e. PQ = QP). PERMUTATION MATRICES 245

Proof. Since P, Q are symmetric and Rt = R- ’ for any R E ff,, where t means transpose, we have by Lemma 1

P and Q are complementary t) 1, PQ are complementary

* PQ is symmetric

0 PQ=QP. n

We now introduce another concept.

DEFINITION 2. Let X c IIn and p be a prime number. Then X is called a p-set if Pp = I, for all P E X, and a p-set is called a commute p-set if PQ=QP forall P,QEX.

From Lemma 1, Lemma 4, and Definition 2, we have

COROLLARY 1. Let P,, ., P,,, E lI,, be symmetric and pairtiise comple- mentary. Then (PI,. , P,,,} is a commute e-set.

COROLI,AKY2. Let P,, P,, . . , P,,, E n,, be pairwise complementary. Then {I,, PF’P,, . , PI’P,,,} is a commute g-set.

We shall first consider the structure of a commute p-set. Let G be a subgroup of S,,, the symmetric group of degree n. Then G acts on the set N = (1,. . , n] in the usual way. For any x E N, we shall denote by G(x) the orbit of x under G [i.e. G(x)= (v(x):u E G}]. For any finite set X, we shall denote by 1x1 the number of elements in X. Noting that the orbits form a partition of N, we easily obtain the following lemma.

LEMMA 5. Let G be a subgroup of S,, r E S,,, and x,x1. y, y1 E N such that x1 E G(x), y1 E G(y), T(x~) = yl, and TQ = UT for all v E G. Then

(i) r(G(x)) = G(y) and consequently IG(x)l = IG(y)l; (ii) the restriction 7jcC,) of T on G(x) is uniquely determined by x, and yl.

Let u E S,. Define P, = (pij)i,j=l,,,,,n, where pjj = aiaCjj and aik is the Kronecker delta. Then P, E n,, and the correspondence from u to P, is an 246 YIK-HOI AU-YEUNG AND CHE-MAN CHENG

isomorphism from S, onto II,. If x E N and 9 C II,, then we define P,(r) = o(x) and 9(x) = {P(X): P E 9). For any two square matrices A and B, not necessarily of the same size, we denote by A@ B and A@ B the direct sum and tensor product respectively.

COROLLARY 3. Let 9 be a subgroup of n,, such that

&(;$,+1) ={ ~EJi+l,~~&i+,,., ;&+n*} for s=l,...,Z,

where ni, i = 1,. ,I, are positive integers, and Cl = ,ni = n (hence P E 9 * P= P,@ .- . @P,, where Pi E II,C for i = 1, . . , 1). Suppose that Q = (Qsl)s,b=l ,..,, l E II,, where Qst is an n,s X n, matrix such that QP = PQ for all P E 9. Then, for any l< s, t < 1, ifQst + 0, then n, = n,, Q,s, E II,%, and Qsr is uniquely determined by any of its nonzero entries.

Proof. From the definition we see that if P E Kl, and x E N, then P(x) is the position of the (unique) 1 that appears in the x th column. Suppose the (a,P)th element of Q,, is 1, where 1~ LYQ n,s and 1 < j3 < n,. Then from our assumption we have

s-1 s-l Cn,+aES En,+1 , i=l i i=l I

t-1 t-1

Cnj+/3E8 Cnj+l , j=l i j=l I and

Our result then follows from Lemma 5. n

Let p be a prime number. We give two examples of commute p-sets. PERMUTATION MATRICES 247

EXAMPLE 1. Let (T be the cyclic permutation (l,~, p - 1,. . ,2) ( E S,) and C,, = P, ( E n,). For any 1 >/ 1 we define the following matrices in ff,f:

P/“= z 631 c3 . . . @.c I’ I’ /I

Then the set

is a commute p-set in ll),,. The subgroup #trl generated by XLII is also a commute p-set which is given by

d[,]= {Cp3q~ . . . ~CI~:l~i,~pforj=l,...,I}.

Furthermore, for any 1~ i, j < p’, there is a unique P E gtlI such that the (i, j)th element of P is 1. Consequently gt;rI(l) = (1,. . , ~‘1, and, by Gd- lary 3, for any Q E fI,,/, if Q commutes with each element in XIII, then Q E &[;I]. We shall define .9t;ol = i(1)).

EXAMPLE 2. Suppose that li >, 0 for i = 1,. . , y and n = x7= 1~‘1. Then the subgroup (of fI,,) defined by

={PEII,:P= P,@ ... @PC,, PiELgr[l,l for i=l,..., q) 9 [I,.. ./,,I

is a commute p-set. Furthermore,

for l

THEOREM 2. Suppose that 9 is a subgroup of n,, und is a commute p-set. Then there exist li B 0 f&r i = 1,. . . , q and Q E n, such that n = Cp= ,p’l and 9 is a subgroup of the group Qdra,,,,,,,, ,,,,Q-‘. 248 YIK-HOI AU-YEUNG AND CHE-MAN CHENG

Proof. We shall prove the theorem by induction on n. It is obviously true when n = 1. Assume that n > 1 and the theorem holds for all k < n. Suppose that 9 is a subgroup of II, and is a commute p-set. We may assume 191 > 1. Take P E 3 \(I,). Since Pp = I,, there exists Q E II, such that

If n - mp > 0, then by considering the subgroup generated by Q- ’ PQ and using Corollary 3, we see that for any R E 9,

Q-‘RQ = R,@R,, where R I E Il,,,,, and R, E II,,_-,,,,,. Our result then follows from induction. So we may assume p 1n and there is Q E II,, such that

Let I be the largest positive integer such that there exists Q E Il,, with

6 P,"',..., 4 P;‘kQ-‘YQ. (1) 1 1

Then we have two cases:

Case 1. a=l. In this case, from Example 1 we have

Case 2. (Y > 1. In this case, for any R E lIrl, we shall write R = (R,A= ,.. .a’ where R,$, are p’ X p’ matrices. If for any R E Qp ‘SQ we have R,,, = 0 for all s # t, then by case 1, wc have

Q-‘

Rr’ = I,, we can conclude that

where oi E HP, and ap’ = n = BP’+’ + ypl with p > I. Now we write

(3) for j = l,..., 1. From (2) and (3) and Corollary 3, we see that if S E Q- ‘SQ, then S = S,@S,, where S, E Ilar~+l an d S, E IIv,,/. If y > 0, then our result follows from induction assumption. If y = 0, then from (2) and (3) and the definition of 1, we see that this case is impossible. Hence Theorem 2 is completely proved. n

COROLLARY 4. Sumrose that X is a commute p-set in II... Then there exist li > 0 for i = 1,. , q &l Q E II,1 such that n = k:/=, ~‘1 arh

Back to our problem. When p = 2, we have

a b conv 9t,] = :a,I?>O, a+b=l , (( 17 a 1 1

/a b c d\ h a d c a,b,c,d>O conv zYzl = c d a b ‘a+b+c+d=l ’ i. d c b a/ I

‘(a b c d e f g hi b adcfehg cdabghef dcbahgfe a,...,h>O conv331=( e f g h a b c d ‘a+ ... +h=l fehgbadc g h e f c d a b \\h g f e d c b al 250 YIK-HOI AU-YEUNG AND CHE-MAN CHENG

For 1= 1,2,3, every matrix in conv ~9~~~is OS. In fact, its corresponding unitary matrix can be taken as

and

respectively.

THEOREM 3. LRtn<15 andP ,,..., P,,,En,. ZfP ,,..., P,,, arepairu&e complementary, then conv( P,, . , P,,,] c en.

Proof. From Lemmas 1 and 2, we may assume P, (i = 1,. , m> are symmetric. Then, by Corollary 1, {P,, . . , P,,,} is a commute 2-set, and by Corollary 4, there exist Zi > 0 for i = 1,. . . , q such that n = Cj’= 121a and

Since n < 15, we see that li < 3 for i = 1,. . . , q, and consequently conv+ I3. . ..lyl C en. Hence conv{ P,, . . , P,,} C en. n PERMUTATION MATRICES 251

3. SOME REMARKS

REMARK 1. Making use of Corollary 4, we can show that the maximum number of permutation matrices such that conv{P,} c @n is 2tn/‘I, where [n /2] denotes the integral part of n /2. An example of such permutation matrices is given by (with p = 2)

9 (l,l.l,..., 11 if n is even and

if n isodd. 9 [l,l.l, ..,l,Ol

REYAHK 2. Write #t;I1 = {Pt};?l 1 with P, = IZ/. The construction of or- thogonal matrices for convgt;Il, 1 = 1,2,3, is as follows:

We assign a real number /3 with ]pI = 1 to each of the nonzero entries of Pi in such a way that the two columns of each of the “2X2 complementay block’ of any pair of P, are perpendicular. If we write the assigned Pi as Pi, then an orthogonal matrix for CcriPi is given by C&pi.

For general 1, if such an assignment of /3 is possible, we may assume P, = Pi = I?I. Then it can be shown that (pJ:L, must satisfy the Hurwitz matrix equations given by

B,, + BI, = 0,

B,“=-I, (4)

B,, B, + B, B, = 0 for h+k.

It is known (for example, see [7]) that the maximum number of matrices satisfying the Hurwitz matrix equations is p(n) - 1, where p(n) denotes the Hurwitz-Radon number defined as follows: if

n = (2~ + 1)2” and b=c+4d, O

p(n)=2’+8d.

As ~(16) = 9, the method fails for 2 > 4. In the construction if we take complex number /3 instead of real, then {~J,$ also satisfy the equations (4) 252 YIK-HOI AU-YEUNG AND CHE-MAN CHENG except that we replace “transpose” by “transpose conjugate” in the first equation. Consequently, when n = 16 (for example, see [l] or [3]), the number of such matrices cannot exceed 9. So again the method fails. We do not know whether Theorem 3 holds or not for n > 15. Our construction of the unitary matrix is too simple in the sense that the pi’s do not depend on the convex coefficients cri. It may be interesting to note that while the 16X 16 matrix below is orthostochastic, our construction fails to show it:

iii; 1 ..* 1 \ 7%

1 1 ii; ... TG)

REMARK 3. Let 9 be a subgroup of II,. If 9 is also a 2-set, then 9 will automatically be abelian. However, for prime p > 3, this implication holds if and only if n < p”. This can be seen by the following proposition:

PROPOSITION 1. L.et p he an odd prime.

(i) lf n < p” and 9 is a subgroup of EI, which is also a p-set, then 9 is abelian. (ii) rfn = p’, then th ere exists a nonabelian subgroup 9 of TI,, such that 9 is also a p-set.

Proof. (i): Write n = lp + d, where 0 < 1< p and 0 < d < p. Similar to the proof of Theorem 2, we can show that there exists Q E II, such that

which is an abelian group. (ii): The subgroup 9 may be taken to be the group generated by C,,@I,) n and ZP @C P @C’@I’ ... @C”-‘.I’

REFERENCES

1 Y. H. Au-Yeung, On matrices whose nontrivial real linear combinations are nonsingular, Proc. Amer. Math. Sot. 29:17-22 (1971). 2 Y. H. Au-Yeung and Y. T. Poon, 3 X 3 orthostochastic matrices and the convexity of generalized numerical ranges, Linear AZgebru Appl. 27:69-79 (1979). 3 S. Friedland, J. W. Robbin, and J. H. Sylvester, On the crossing rule, Comm. Pure Appl. Math. 37:19-37 (1984). PERMUTATION MATRICES 253

4 A Horn, Doubly stochastic matrices and the diagonal of a rotation matrix, Amer. J. Math. 76:620-630 (1954). 5 M. Marcus, K. Kidman, and M. Sandy, Products of elementary doubly stochastic matrices, Linear and Multilinear Algebra 15:331-340 (1984). 6 Y. T. Poon and N. K. Tsing, Inclusion relations between orthostochastic matrices and products of pinching matrices, Linear and Multilinear Algebra 21:253-259 (1987). 7 Y. C. Wong, lsoclinic n-Planes in Euclidean 2n-Space, Cliffwd Parallels in Elliptic (2n - l&Space, and the Hurwitz Matrix Equation, Mem. Amer. Math. Sot. 41, 1961; second printing with corrections and minor changes, 1971.

Received 13 September 1989; final manuxript accepted 20 ]une 1990